About

LLM Compass is a knowledge compass for LLM training algorithms — from base models, SFT, LoRA, the DPO family and PPO·GRPO family, to distillation, inference & decoding, and Harness / Agent / Skills. Each method aims to give a one-line definition, the core formula, a diagram, and practical tuning notes, with the original paper and year of proposal.

Curation principle: only well-known, widely-discussed algorithms — no obscure variants, no quickly-stale benchmark numbers; qualitative conclusions first, exact figures per the original papers.

Author

Maintained by Brench. Feel free to reach out, suggest corrections, or contribute:

• Xiaohongshu (RED): profile
• GitHub: @BrenchCC
• WeChat: BrenchChen (scan the QR below, or add me directly — please mention why you're reaching out)

Contributing

• Repository: BrenchCC/llm-compass — issues and PRs welcome.

License

• Code: MIT
• Content: CC BY-SA 4.0

Original paper figures referenced on this site remain the copyright of their authors and are used for educational annotation only.